A video sequence consists of a number of pictures, usually called frames. Subsequent frames are very similar, thus containing a lot of redundancy from one frame to the next. Before being efficiently transmitted over a channel or stored in memory, video data is compressed to conserve both bandwidth and memory. The goal is to remove the redundancy to gain better compression ratios. A first video compression approach is to subtract a reference frame from a given frame to generate a relative difference. A compressed frame contains less information than the reference frame. The relative difference can be encoded at a lower bit-rate with the same quality. The decoder reconstructs the original frame by adding the relative difference to the reference frame.
A more sophisticated approach is to approximate the motion of the whole scene and the objects of a video sequence. The motion is described by parameters that are encoded in the bit-stream. Pixels of the predicted frame are approximated by appropriately translated pixels of the reference frame. This approach provides an improved predictive ability over a simple subtraction. However, the bit-rate occupied by the parameters of the motion model must not become too large.
In general, video compression is performed according to many standards, including one or more standards for audio and video compression from the Moving Picture Experts Group (MPEG), such as MPEG-1, MPEG-2, and MPEG-4. Additional enhancements have been made as part of the MPEG-4 part 10 standard, also referred to as H.264, or AVC (Advanced Video Coding). Under the MPEG standards, video data is first encoded (e.g. compressed) and then stored in an encoder buffer on an encoder side of a video system. Later, the encoded data is transmitted to a decoder side of the video system, where it is stored in a decoder buffer, before being decoded so that the corresponding pictures can be viewed.
The intent of the H.264/AVC project was to develop a standard capable of providing good video quality at bit rates that are substantially lower than what previous standards would need (e.g. MPEG-2, H.263, or MPEG-4 Part 2). Furthermore, it was desired to make these improvements without such a large increase in complexity that the design is impractical to implement. An additional goal was to make these changes in a flexible way that would allow the standard to be applied to a wide variety of applications such that it could be used for both low and high bit rates and low and high resolution video. Another objective was that it would work well on a very wide variety of networks and systems.
H.264/AVC/MPEG-4 Part 10 contains many new features that allow it to compress video much more effectively than older standards and to provide more flexibility for application to a wide variety of network environments. Some key features include multi-picture motion compensation using previously-encoded pictures as references, variable block-size motion compensation (VBSMC) with block sizes as large as 16 pixels×16 pixels and as small as 4 pixels×4 pixels, six-tap filtering for derivation of half-pel luma sample predictions, macroblock pair structure, quarter-pixel precision for motion compensation, weighted prediction, an in-loop deblocking filter, an exact-match integer 4 pixel×4 pixel spatial block transform, a secondary Hadamard transform performed on “DC” coefficients of the primary spatial transform wherein the Hadamard transform is similar to a fast Fourier transform, spatial prediction from the edges of neighboring blocks for “intra” coding, context-adaptive binary arithmetic coding (CABAC), context-adaptive variable-length coding (CAVLC), a simple and highly-structured variable length coding (VLC) technique for many of the syntax elements not coded by CABAC or CAVLC, referred to as Exponential-Golomb coding, a network abstraction layer (NAL) definition, switching slices, flexible macroblock ordering, redundant slices (RS), supplemental enhancement information (SEI) and video usability information (VUI), auxiliary pictures, frame numbering and picture order count. These techniques, and several others, allow H.264 to perform significantly better than prior standards, and under more circumstances and in more environments. H.264 usually performs better than MPEG-2 video by obtaining the same quality at half of the bit rate or even less.
MPEG is used for the generic coding of moving pictures and associated audio and creates a compressed video bit-stream made up of a series of three types of encoded data frames. The three types of data frames are an intra frame (called an I-frame or I-picture), a bi-directional predicated frame (called a B-frame or B-picture), and a forward predicted frame (called a P-frame or P-picture). These three types of frames can be arranged in a specified order called the GOP (Group Of Pictures) structure. I-frames contain all the information needed to reconstruct a picture. The I-frame is encoded as a normal image without motion compensation. On the other hand, P-frames use information from previous frames and B-frames use information from previous frames, a subsequent frame, or both to reconstruct a picture. Specifically, P-frames are predicted from a preceding I-frame or the immediately preceding P-frame.
Frames can also be predicted from the immediate subsequent frame. In order for the subsequent frame to be utilized in this way, the subsequent frame must be encoded before the predicted frame. Thus, the encoding order does not necessarily match the real frame order. Such frames are usually predicted from two directions, for example from the I- or P-frames that immediately precede or the P-frame that immediately follows the predicted frame. These bidirectionally predicted frames are called B-frames.
There are many possible GOP structures. A common GOP structure is 15 frames long, and has the sequence I_BB_P_BB_P_BB_P_BB_P_BB_. A similar 12-frame sequence is also common. I-frames encode for spatial redundancy, P and B-frames for both temporal redundancy and spatial redundancy. Because adjacent frames in a video stream are often well-correlated, P-frames and B-frames are only a small percentage of the size of I-frames. However, there is a trade-off between the size to which a frame can be compressed versus the processing time and resources required to encode such a compressed frame. The ratio of I, P and B-frames in the GOP structure is determined by the nature of the video stream and the bandwidth constraints on the output stream, although encoding time may also be an issue. This is particularly true in live transmission and in real-time environments with limited computing resources, as a stream containing many B-frames can take much longer to encode than an I-frame-only file.
B-frames and P-frames require fewer bits to store picture data, generally containing difference bits for the difference between the current frame and a previous frame, subsequent frame, or both. B-frames and P-frames are thus used to reduce redundancy information contained across frames. In operation, a decoder receives an encoded B-frame or encoded P-frame and uses a previous or subsequent frame to reconstruct the original frame. This process is much easier and produces smoother scene transitions when sequential frames are substantially similar, since the difference in the frames is small.
Each video image is separated into one luminance (Y) and two chrominance channels (also called color difference signals Cb and Cr). Blocks of the luminance and chrominance arrays are organized into “macroblocks,” which are the basic unit of coding within a frame.
In the case of I-frames, the actual image data is passed through an encoding process. However, P-frames and B-frames are first subjected to a process of “motion compensation.” Motion compensation is a way of describing the difference between consecutive frames in terms of where each macroblock of the former frame has moved. Such a technique is often employed to reduce temporal redundancy of a video sequence for video compression. Each macroblock in the P-frame or B-frame is associated with an area in the previous or next image that it is well-correlated, as selected by the encoder using a “motion vector.” The motion vector that maps the macroblock to its correlated area is encoded, and then the difference between the two areas is passed through the encoding process.
Conventional video codecs use motion compensated prediction to efficiently encode a raw input video stream. The macroblock in the current frame is predicted from a displaced macroblock in the previous frame. The difference between the original macroblock and its prediction is compressed and transmitted along with the displacement (motion) vectors. This technique is referred to as inter-coding, which is the approach used in the MPEG standards.
One of the most time-consuming components within the encoding process is motion estimation. Motion estimation is utilized to reduce the bit rate of video signals by implementing motion compensated prediction in combination with transform coding of the prediction error. Motion estimation-related aliasing is not able to be avoided by using inter-pixel motion estimation, and the aliasing deteriorates the prediction efficiency. In order to solve the deterioration problem, half-pixel interpolation and quarter-pixel interpolation are adapted for reducing the impact of aliasing. To estimate a motion vector with quarter-pixel accuracy, a three step search is generally used. In the first step, motion estimation is applied within a specified search range to each integer pixel to find the best match. Then, in the second step, eight half-pixel points around the selected integer-pixel motion vector are examined to find the best half-pixel matching point. Finally, in the third step, eight quarter-pixel points around the selected half-pixel motion vector are examined, and the best matching point is selected as the final motion vector. Considering the complexity of the motion estimation, the integer-pixel motion estimation takes a major portion of motion estimation if a full-search is used for integer-pixel motion estimation. However, if a fast integer motion estimation algorithm is utilized, an integer-pixel motion vector is able to be found by examining less than ten search points. As a consequence, the computation complexity of searching the half-pixel motion vector and quarter-pixel motion vector becomes dominant.
Due to the advance of semiconductor technology that reduces the size of a transistor continuously so that it increases the silicon space for implementing more functionality, and the physical limitations that cause significant heat dissipation, processor manufacturers have started to introduce multi-core or multi-processor systems. Video compression processing demands processor power and attempts to exploit the computational power that contemporary processors are able to provide. However, many video compression methodologies are either sequentially-defined or there exists dependence between neighborhood video objects so that it is difficult to exploit the thread level parallelism for video compression processes on contemporary parallel computing systems. One of the approaches to exploit the parallel processing power in video compression applications is to split a picture in video sequence to multiple parts horizontally, known as slices. Since there is no semantical dependency between slices of multi-slice, this is suitable for video compression in multi-processing systems.
However, some video compression applications require a single slice approach (one slice per picture). With the single slice approach, there are many dependency issues in single slice syntax and semantics around block boundary, especially in the Advanced Video Coding (AVC) specification.
One method of parallel processing video compression on a multi-processing system with the single slice approach is to divide a picture horizontally to top half and bottom half. One thread processes the top half of the picture and the other thread processes the bottom half. Both threads process the same picture. The bottom thread ignores dependency around the boundary and also handles conformance of syntax around the boundary. When the bottom thread processes the first Macroblock (MB) lines, it selects the MB mode that is independent of its upper MB. However, this kind of general methodology causes worse efficiency of compression than the standard single slice raster scan approach.
Another parallel processing methodology of video compression is to split a picture into multiple parts horizontally. These horizontally separated parts are called slices. The video compression system then encodes these slices in a parallel processing way. This method produces a multiple slice result.
However, the multiple slice approach suffers many problems. The encoded result of multiple slices is different from single slice because the encoded result of multiple slices is semantically independent of each other. Therefore, it is difficult or impossible to validate the correctness of parallel processing methodology by multi-slices. The video quality decreases at the boundaries of slices. Video compression using horizontal multi-slices encoding suffers workload unbalance if the complexity of video contents are different in different slices. The result of individual slices of horizontal multi-slices encoding needs to be concatenated to form a single result. This is additional work that does not exist in single slice encoding.